I'm Stewart, an AI engineer building practical tools around local models, coding agents, evaluation, and accessibility.
Before working independently, I spent more than 10 years at Microsoft as an Escalation Engineer working on complex Exchange Server issues. That experience shapes how I build: reproduce the problem, understand the failure, and verify the fix.
Blog and experiments · LinkedIn
Open-source contributions
I contribute to oMLX, an inference server for Apple Silicon. My merged contributions include:
- Model-owned batch cache conversion, fixing how custom model caches enter batched inference.
- Batching under scheduler contention, adjusting scheduling when multiple requests compete with speculative decoding.
- DeepSeek V4 prefill optimization, skipping redundant indexer scoring when every pooled row already fits.
Selected projects
Ground Ball
A local query engine for historical MLB data. Natural-language questions and structured recipes compile into validated query plans and execute against DuckDB. Results include the exact SQL, parameters, and source fingerprints. The primary query path is deterministic and runs offline without an LLM.
PTT Dictation
A Windows dictation app with local speech recognition and live text insertion in supported applications. Built with C# and .NET, with Parakeet models behind a replaceable transcription interface and a CPU fallback.
ObviousSaveMD
A native macOS Markdown editor built with SwiftUI and AppKit. Plain-text files, conservative file loading, and tests for document state and save behavior.
My current work centers on local inference performance, model evaluation, reliable agent workflows, and tools that make computers easier to use.